JVET-Q0100 On Maximum Block Size for Chroma Transform Skip [T. Tsukuba, M. Ikeda, Y. Yagasaki, T. Suzuki (Sony)]
In VVC WD, there is an inconsistency between maximum transform block size for chroma and transform skipped block size for chroma; Maximum transform skipped block size for chroma is always set to (1 << ( log2_transform_skip_max_size_minus2 + 2 ) regardless of chroma format while maximum transform block size for chroma is set depending on chroma format. This contribution proposes to derive maximum block size for chroma transform skip based on log2_transform_skip_max_size_minus2, SubWidthC, and SubHeightC.
At normal QP condition, it is reported that proposal gives average BD-rate differences of (0.00%, -0.01%, -0.01%) for AI, (0.00%, 0.02%, 0.06%) for RA and (0.02%, 0.15%, -0.06%) for LB, respectively.
At low QP condition (QP=2, 7, 12, 17), it is reported that proposal gives average BD-rate differences of (0.00%, 0.00%, 0.01%) for AI, (0.00%, 0.00%, 0.00%) for RA and (0.00%, 0.02%, 0.00%) for LB, respectively. Especially, for SCC sequence, it is reported that proposal provides some BD-rate saving: (-0.14%, -0.13%, -0.11%) for AI, (-0.06%, 0.24%, 0.24%) for RA and (-0.39%, -0.25%, -0.26%) for LB.
It is pointed out that the proposal, though appearing logical in terms of coupling luma and chroma sampling also in transform skip, would somehow restrict flexibility of encoder choices. It is a design intent that luma and chroma can select TS independently of each other. For example, luma could use 64x64 regular transform, and chroma 32x32 TS (which would no longer be possible if the proposal were adopted). It is also not fully clear how it would affect the dual-tree case.
No action was taken on this.